Senior Systems Software Engineer, AI Stack and Performance - DGX Station

Details

DGX Station (Galaxy) is NVIDIA’s workstation-class AI computer—built on GB300 Blackwell GPUs with NVLink interconnect, delivering data-center-grade AI compute in a deskside form factor. DGX Station is shipped to OEM and OSV partners as a complete software/firmware GA release including firmware bundles, DGX BaseOS, GPU drivers, CUDA toolkit, DCGM, and DOCA/OFED. The role owns AI stack readiness on DGX Station: profiling workloads, identifying bottlenecks across GPU compute, NVLink, memory, and host interconnects, driving optimizations across the full stack—from GPU kernels through frameworks to applications—and working hands-on with framework, compiler, and GPU architecture teams to ensure DGX Station delivers best-in-class performance for real AI workloads in multi-user and multi-GPU configurations.

Responsibilities

• Own production readiness of AI applications on DGX Station (NemoClaw, Hermes agents, NIM microservices, and key customer workloads). Define "ready to ship" criteria, run validation, and close gaps between "it runs" and "it runs well" across single-GPU and multi-GPU configurations.
• Profile and optimize LLM and deep learning workloads (PyTorch, TensorFlow, JAX) across training and inference on GB300 Blackwell multi-GPU architecture. Characterize performance across model sizes, batch sizes, precision modes (FP16, INT8, FP8), and GPU scaling (single-GPU vs. multi-GPU with NVLink).
• Identify system-level bottlenecks in GPU compute, NVLink bandwidth, host memory, PCIe, and CPU–GPU communication. Implement or drive optimizations across the stack: kernel tuning, memory placement, NVLink utilization, data pipeline efficiency, and scheduling to increase throughput on DGX Station’s multi-GPU topology.
• Collaborate with framework, compiler (TensorRT, NVCC, Triton), and GPU architecture teams to improve kernel fusion, graph execution, operator scheduling, and memory management for Blackwell GPUs. Translate platform-specific constraints and multi-GPU topology into actionable optimization requests for upstream teams.
• Validate multi-user and concurrent workload scenarios—multiple users running simultaneous training jobs, inference serving alongside development, and resource isolation via MIG or time-slicing—to ensure reliable shared-workstation performance.
• Validate the full NVIDIA AI software stack on DGX Station: CUDA toolkit, cuDNN, TensorRT, NCCL, Triton Inference Server, DCGM, and DOCA/OFED; ensure version compatibility, functional correctness, and performance parity with reference data center configurations.
• Build and maintain performance benchmarking infrastructure for DGX Station—automated regression tracking across key models (LLaMA, GPT, Stable Diffusion, Whisper), framework versions, and driver updates. Make performance data visible and actionable for GA release decisions.
• Work with product management and OEM/OSV partners to understand target use cases (local LLM training and inference, agentic AI, multi-user research, RTX Pro workloads) and ensure DGX Station delivers compelling performance for each. Support customer deployment readiness and field critical issues.

Requirements

• BS or MS (or equivalent experience) in Computer Science, Electrical Engineering, or related field.
• 12+ years in systems software engineering with hands-on experience in AI/ML workload optimization, GPU performance analysis, or deep learning infrastructure.
• Strong proficiency with deep learning frameworks (PyTorch, TensorFlow, or JAX), including internals: graph execution, operator dispatch, memory management, and custom kernel integration.
• Experience profiling and optimizing GPU workloads using Nsight Systems, Nsight Compute, CUPTI, or equivalent. Ability to read GPU traces and translate observations into actionable optimizations.
• Strong understanding of GPU architecture: compute units, memory hierarchy, NVLink, multi-GPU scaling, and how they impact AI workload performance.
• Experience with inference optimization: quantization (INT8/FP8), model compilation (TensorRT, torch.compile), batching strategies, and serving frameworks.
• Proficiency in C/C++, CUDA, and Python. Comfortable reading and modifying GPU kernels.

Ways to stand out from the crowd

• Experience optimizing LLM training or inference on multi-GPU NVIDIA systems (DGX, HGX, or multi-GPU workstations).
• Contributions to open-source AI frameworks, CUDA libraries, or inference engines.
• Experience with multi-GPU communication optimization—NCCL tuning, NVLink utilization, collective operations, and parallel training strategies.
• Track record of collaborating with compiler and hardware architecture teams to drive kernel fusion, graph optimization, or hardware-specific performance improvements.
• Experience shipping AI-powered products where application performance on specific hardware was a hard shipping requirement.

Benefits

• Base salary range: 224,000 USD - 356,500 USD (determined based on location, experience, and pay of employees in similar positions).
• Eligible for equity and NVIDIA benefits (link provided in original posting).

Additional information

• Applications accepted at least until June 5, 2026. This posting is for an existing vacancy.
• NVIDIA uses AI tools in its recruiting processes and is an equal opportunity employer committed to fostering an inclusive work environment.